ARCHE Lets Chemistry Agents Iteratively Eliminate Mechanistic Hypotheses Using Computational Results, but the Public Release Cannot Fully Reproduce the Paper

The research team released ARCHE, a system that goes beyond having agents generate textual descriptions of chemical reaction mechanisms. Instead, it submits candidate explanations to real computational workflows and revises its conclusions based on evidence from energies, structures, and reaction pathways. The system first retrieves literature, then generates, merges, and ranks hypotheses from multiple research perspectives before selecting five for separate investigation. For each hypothesis, the MiniChem toolkit handles molecular preparation, electronic structure calculations, reaction pathway analysis, and result interpretation. Finally, Task Review—which can only read evidence and cannot submit calculations itself—decides whether to request additional experiments, move to the next batch of hypotheses, or answer the question.

The paper demonstrates three types of tasks: reconstructing stereoselective transition states for an asymmetric reaction; investigating the C–I bond-cleavage pathway of an α-iodoboronate ester under 450 nm irradiation; and identifying a torsional strain-energy descriptor that explains selectivity in nickel-catalyzed migratory cross-coupling. In the first case, the reported backbone RMSDs for the major and minor transition states were 0.01 and 0.15 Å, respectively, with a calculated activation free-energy difference of approximately 1.9 kcal/mol. In the second case, the system obtained an excitation energy of about 64.0 kcal/mol, close to the 63.6 kcal/mol energy of a 450 nm photon.

The public Arche-Harness also adds SQLite-backed working memory, isolated investigation sessions, a read-only review interface, and `--resume-task`, allowing long-running quantum chemistry jobs to continue after interruption. Its tool catalog registers 65 actions and 28 backends, but actual execution still depends on Gaussian, Multiwfn, model APIs, and substantial computational resources. The default service configuration even allocates 60 CPU cores, 320 GB of memory, and one GPU, making a complete deployment impractical on a typical workstation.

There are also clear limits to what has been made public. The paper uses ARCHE-Chem, trained from Qwen2.5-7B-Instruct, along with a reward model, but their weights, training corpus, and training pipeline have not been released. The control flow in the current harness also differs from the version described in the paper. The project’s own reanalysis scripts further flag discrepancies in the number of candidate records and energy values for Case 1, a different complex-formation free energy for Case 2, and the absence of the reported Wiberg bond order for Case 3. As a result, this release is better suited to auditing the architecture and existing data than to independently reproducing the three chemical conclusions. The next critical step is to release the models, original computational inputs, and reaction-pathway validation data.